Real-time traceability method of width of defect based on divide-and-conquer

Abstract

In a real-time traceability method of a width of a defect based on divide-and-conquer provided by the present invention, through the calibration transfer function, the multidimensional eigenvector analysis technology based on the electromagnetic field simulation database of defect scattered dark-field imaging and the adaptive threshold segmentation method, the real-time traceability of the width of the defect greater than and close to the diffraction limit of the system is performed, respectively. The extreme random tree regression model is trained by multidimensional eigenvector analysis technology based on the multidimensional eigenvectors in the electromagnetic field simulation database of the defect scattered dark-field imaging. The present invention solves the problems that the width of the defect in defect detection is difficult to be accurately measured in real time, and the conventional image processing algorithm is difficult to accurately identify the width of the defect close to the diffraction limit of the system.

Description

CROSS REFERENCE OF RELATED APPLICATION

This is a U.S. National Stage under 35 U.S.C 371 of the International Application PCT/CN2019/099008, filed Aug. 2, 2019, which claims priority under 35 U.S.C. 119(a-d) to CN 201910516180.1, filed Jun. 14, 2019.

BACKGROUND OF THE PRESENT INVENTION

Field of Invention

The present invention relates to the field of machine vision inspection technology, and more particularly to a real-time traceability method of a width of a defect based on divide-and-conquer.

Description of Related Arts

With the advancement of industrial automation manufacturing and international high-precision technology research, more and more stringent requirements are put forward for the surface quality of precision components. Especially in the field of high-energy laser systems, the existence of defects on the surface of smooth glass elements may cause the elements to burst and cause the system to fail to operate. Among the defects, the surface defects of large-diameter components, due to the small size and random distribution relative to the component diameter, are the difficulty of surface quality control. A very important evaluation index for defects is the width of the defect. In order to overcome the problems of low efficiency, strong subjectivity, and ambiguity in traditional manual visual inspection, combining machine vision inspection technology and pattern recognition technology, using scattered dark-field imaging technology to obtain digital images of micron-level defects on the surface of large-diameter components; aiming at the problem that it is difficult for conventional image processing algorithms to accurately measure the width of the defect in real time, the calibration transfer function is used to trace the width of the defect in real time; aiming at the problem that the calibration transfer function is difficult to accurately trace the width of the defect when the width of the defect is close to the diffraction limit of the system, the multidimensional eigenvector in the electromagnetic field simulation database of the defect scattered dark-field imaging is used as the data basis, the width of the defect which is close to the diffraction limit is traced through training an extreme random tree regression model; for defects with a smaller width, an adaptive threshold segmentation algorithm is used for traceability. Combining the above three situations, a real-time traceability method of a width of a defect based on divide-and-conquer with excellent portability is proposed to objectively detect and evaluate the width of the defect.

SUMMARY OF THE PRESENT INVENTION

An object of the present invention is to overcome the shortcomings of the prior art and provide a real-time traceability method of a width of a defect in the microscopic scattered dark-field imaging detection environment, so as to improve the recognition accuracy of the width of the defect.

To solve the above technical problem, the present invention provides a technical solution as follows.

A real-time traceability method of a width of a defect based on divide-and-conquer comprises steps of:

(1) estimating the width of the defect through a calibration transfer function (CTF) based on standard defect digital calibration technology, recording the width of the defect estimated through the CTF as W_CTF, wherein if W_CTF is greater than 5 μm, which means that the width of the defect complies with geometric imaging law and is located in a linear segment of the CTF, then W_CTF acts as an accurately estimated value of an actual width of the defect; if W_CTF is less than or equal to 5 μm, which means that the width of the defect is smaller, diffraction broadening occurs during imaging, the width of the defect no longer complies with the geometric imaging law, and W_CTF acting as the estimated value of the actual width of the defect has a large error, then at this time, a step of (2) is executed;

(2) estimating the width of the defect through a multidimensional eigenvector analysis technology based on an electromagnetic field simulation database of defect scattered dark-field imaging, recording the width of the defect estimated through the multidimensional eigenvector analysis technology as a regression width W_regrs, wherein if W_regrs is greater than a diffraction limit σ₀ of a dark-field imaging system, then the regression width W_regrs acts as the accurately estimated value of the actual width of the defect; if W_regrs is less than or equal to the diffraction limit σ₀, the regression width W_regrs is unable to distinguish different widths in a width range, that is, less than or equal to the diffraction limit σ₀, then at this time, a step of (3) is executed; and

(3) processing a dark-field image and extracting the width of the defect through an adaptive threshold segmentation method, recording the width of the defect extracted through the adaptive threshold segmentation method as an imaging width W_img, wherein if the imaging width W_img is greater than the diffraction limit σ₀, the imaging width W_img is unable to distinguish different widths, then the regression width W_regrs still acts as the estimated value of the actual width of the defect; if the imaging width W_img is less than or equal to the diffraction limit co, then the imaging width W_img acts as the estimated value of the actual width of the defect.

In the step of (1), the width of the defect is estimated through the CTF based on the standard defect digital calibration technology, and the CTF for describing a relationship between the actual width of the defect and a pixel width of the defect after imaging is obtained through steps of:

(1-1) according to a principle of binary optics, etching a standard defect on a fused silica glass substrate through electron beam exposure and ion beam etching technology, and obtaining a standard plate;

(1-2) measuring an actual width of the standard defect on the standard plate with a scanning electron microscope (SEM);

(1-3) collecting a dark-field image of the standard defect on the standard plate with a dark-field imaging system;

(1-4) performing a series of processing such as threshold segmentation and morphological operations on the dark-field image, and obtaining a pixel width of the standard defect; and

(1-5) taking the pixel width of the standard defect as an independent variable, and the actual width of the standard defect measured by the SEM as a dependent variable, and fitting the independent variable and the dependent variable through a least square method, so as to obtain an object-image correlation function which is the calibration transfer function.

From the object-image relationship of geometric optics, it is able to be seen that the pixel width of the defect has a linear relationship with the actual width of the defect. Therefore, the width traceability of most defects is able to be achieved through the CTF. However, with the decrease of the width of the defect, when the width of the defect approaches the diffraction limit of the dark-field imaging system, geometric optical imaging theory is no longer applicable at this time, fine defects will cause diffraction effects, the pixel width will be increased, which will cause while using the CTF for width traceability, a same pixel width corresponds to more than one actual width, so that it is confused and unable to accurately obtain the actual width. Experiments have found that when the width of the defect is less than or equal to 5 μm, the CTF no longer satisfies the linear relationship, and the width traceability is confused. Therefore, the traceability of the width which is less than or equal to 5 μm is processed through the steps of (2) and (3).

In the step of (2), through the multidimensional eigenvector analysis technology based on the electromagnetic field simulation database of defect scattered dark-field imaging, the regression width W_regrs is obtained, which specifically comprises:

(2-1) building the electromagnetic field simulation database of defect scattered dark-field imaging, and obtaining simulation data of scattered light intensity distribution of multiple defects;

(2-2) collecting a dark-field image of a defect to be tested with the dark-field imaging system, obtaining a gray distribution along a width direction of the standard defect on the standard plate, and obtaining experimental data of the gray distribution of an ideal imaging system after deconvolution;

(2-3) performing coordinate system unity and Min-Max normalization on the simulation data obtained by the step of (2-1) and the experimental data obtained by the step of (2-2), and obtaining normalized simulation data of scattered light intensity distribution and normalized experimental data of the gray distribution, wherein the step of (2-3) specifically comprises:

(2-3-1) taking a physical size x as an abscissa for the simulation data obtained by the step of (2-1) and a pixel p as an abscissa for the experimental data obtained by the step of (2-2), unifying the simulation data obtained by the step of (2-1) and the experimental data obtained by the step of (2-2) to take the physical size x as the abscissa through formulas of

$\begin{array}{cc}p=\frac{\mathrm{Mx}}{s_p},& \left(I\right)\end{array}$ $\begin{array}{cc}{f}_{\mathrm{res}}\left(p\right)\stackrel{p=\mathrm{Mx}/s_p}{⟶}{f}_{\mathrm{res}}\left(x\right),& \left(\mathrm{II}\right)\end{array}$
here, a pixel size of a charge-coupled device (CCD) of the dark-field imaging system is expressed by Sp, a magnification of the dark-field imaging system is expressed by A; and

(2-3-2) performing Min-Max normalization on the simulation data obtained by the step of (2-1) and the experimental data obtained by the step of (2-2) by a formula of

$\begin{array}{cc}\{\begin{array}{c}{I}_{\mathrm{norm}}\left(x\right)=\frac{I_{\mathrm{sct}}\left(x\right)-\min \left\{I_{\mathrm{sct}}\left(x\right)\right\}}{\max \left\{I_{\mathrm{sct}}\left(x\right)\right\}-\min \left\{I_{\mathrm{sct}}\left(x\right)\right\}}\\ f_{\mathrm{norm}}\left(x\right)=\frac{f_{\mathrm{res}}\left(x\right)-\min \left\{f_{\mathrm{res}}\left(x\right)\right\}}{\max \left\{f_{\mathrm{res}}\left(x\right)\right\}-\min \left\{f_{\mathrm{res}}\left(x\right)\right\}}\end{array};& \left(\mathrm{III}\right)\end{array}$

(2-4) respectively extracting multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution and multidimensional eigenvectors of the normalized experimental data of the gray distribution;

(2-5) taking the multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution in the step of (2-4) as a sample set, training through an extreme random tree regression algorithm with supervised learning in machine learning theory, and obtaining a regression model of a width of any defect; and

(2-6) taking the multidimensional eigenvectors of the normalized experimental data of the gray distribution in the step of (2-4) as a test set, inputting the regression model obtained by the step of (2-5), estimating the width of the defect to be tested, and obtaining the regression width W_regrs of the defect to be tested.

In the step of (2-1), building the electromagnetic field simulation database of defect scattered dark-field imaging, and obtaining simulation data of scattered light intensity distribution of multiple defects are specifically as follows:

(2-1-1) building an electromagnetic field simulation model of defect scattered dark-field imaging with FDTD (finite-difference time-domain) simulation software;

(2-1-2) building a simulation model of common defects;

(2-1-3) performing numerical calculations of electromagnetic fields with the FDTD simulation software, and obtaining the scattered light intensity distribution of the common defects;

(2-1-4) performing near field—far field projection, obtaining light intensity distribution of the common defects on a far-field image plane of an ideal optical imaging system, and recording as I_sct(x); and

(2-1-5) changing a defect shape and a defect size, continuously repeating the steps of (2-1-3) to (2-1-4), obtaining the scattered light intensity distribution of the defects with different shapes and sizes, obtaining a series of simulation data, wherein the series of simulation data comprise the shape of the defects, the width of the defects and the light intensity distribution I_sct(x) of the defects, so as to form the simulation data of the scattered light intensity distribution of the defects.

In the step of (2-2), collecting the dark-field image of the defect to be tested with the dark-field imaging system, obtaining a gray distribution along a width direction of the standard defect on the standard plate, and obtaining experimental data of the gray distribution of an ideal imaging system after deconvolution are specifically as follows.

(2-2-1) collecting the dark-field image of the defect to be tested with the dark-field imaging system, and extracting the gray distribution ƒ_exp(p) along the width direction of the dark-field image of the defect to be tested;

(2-2-2) measuring a point spread function (PSF) of the dark-field imaging system through a checkerboard diagram, wherein since a PSF of an actual optical system tends to a form of Gaussian function, an expression of the PSF of the dark-field imaging system is obtained through Gaussian function fitting in a measurement process; and

(2-2-3) taking the PSF obtained by the step of (2-2-2) as an initial value, deconvolving the gray distribution ƒ_exp(p) obtained by the step of (2-2-1), and obtaining the gray distribution ƒ_exp(p) of the ideal imaging system after deconvolution.

The step of (2-2-1) that extracting the gray distribution ƒ_exp(p) along the width direction of the dark-field image of the defect to be tested is specifically as follows.

1) performing Otsu threshold segmentation on an original dark-field image, selecting an elongated structural element for opening operations to connect line features of fractures and holes while removing noise, and then extracting long line features which are defect characteristics:

2) performing linear fitting on pixel coordinates (u_i, v_i) of pixels that form the line features through the least square method, and obtaining a straight line expression of av+b=u, wherein both a and h are parameters;

3) obtaining a positive angle γ between the defect and v axis through the parameter a, wherein γ=arctan(a), and rotate the original dark-field image around an upper left corner for −γ, so that the defect is rotated to a vertical direction;

4) calculating a gray histogram of the original dark-field image, wherein a peak value at low gray denotes a gray level of image background, filling a blank area that appears after rotation with the gray of image background, and obtaining a filled image to maintain an image integrity;

• • 5) taking a column average of the filled image as a gray distribution ƒ₀(p) of the defect along the width direction of the defect to be tested, which is unimodal or bimodal, taking a vertical straight line where a peak abscissa p₀ of the unimodal distribution is a symmetry axis, assuming that a data width is 2N+1, and intercepting data points between p₀−N and p₀+N as a gray distribution ƒ_N(p):

6) averaging after making the gray distribution ƒ_N(p) axisymmetric, and obtaining an axisymmetric distribution ƒ_sym(p) which is expressed by a formula of

$\begin{array}{cc}{f}_{\mathrm{sym}}\left(p\right)=\frac{1}{2}\left[f_N\left(p\right)+f_N\left(2p_0-p\right)\right],pϵ\left[p_0-N,p_0+N\right],& \left(\mathrm{IV}\right)\end{array}$
moving coordinate axes, taking an abscissa of the symmetry axis as a coordinate origin, and obtaining a gray distribution ƒ_exp(p) of the dark-field image of the defect along the width direction, which is expressed by a formula of
ƒ_exp(p)=ƒ_sym(p−p₀),p∈[−N,N]  (V).

In the step of (2-4), the multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution and the multidimensional eigenvectors of the normalized experimental data of the gray distribution, respectively comprise curved Gaussian pyramid features of the normalized simulation data and those of the normalized experimental data, an extracting method of the multidimensional eigenvectors comprises:

(2-4-1) recording distributed data as ƒ(n), wherein n represents discrete data points, ƒ_J(n) represents ƒ(n) is located at a J^th floor of a pyramid;

(2-4-2) recording a curve distribution of a j^th floor of the pyramid as ƒ_j(n), performing low-pass Gaussian filtering on ƒ_j(n), and obtaining ƒ′_j(n);

(2-4-3) downsampling ƒ′_j(n) by a factor of 2, and obtaining a curve distribution of a (j−1)^th floor of the pyramid, which is able to be regarded as discarding one data point every one data point, that is
ƒ_j−1(n)=ƒ₂(n)=ƒ_j(2n)  (VI);

(2-4-4) repeating the steps of (2-4-2) and (2-4-3) for P times, wherein j=J, J−1, . . . , J−P+1, and obtaining a P+1 layer Gaussian pyramid; and

(2-4-5) connecting data of the P+1 layer Gaussian pyramid with each other in series, and obtaining the multidimensional eigenvectors.

If the regression width W_regrs is less than or equal to the diffraction limit σ₀, the dark-field image of the defect is degraded seriously and the effect after deconvolution is not good. The grayscale distribution within this width range is very close after normalization, and the regression model is unable to distinguish different widths within this width range. Therefore, the width traceability in this width range is performed through the method described in the step of (3).

The step of (3) comprises processing the dark-field image according to adaptive threshold segmentation method, and obtaining the defect imaging width W_img which is specifically as follows.

(3-1) performing threshold segmentation on the dark-field image through Otsu algorithm;

(3-2) extracting an area A and a length l of the defect, and obtaining a pixel width W_p of the defect, wherein w_p=A/l;

(3-3) through the pixel size S_p of the CCD, and the magnification M of the dark-field imaging system, obtaining a simple imaging width W′_img which is expressed by a formula of

$\begin{array}{cc}{w}_{\mathrm{img}}^{\prime }=\frac{s_p}{M}{w}_p;& \left(\mathrm{VII}\right)\end{array}$
(3-4) performing linear fitting on the simple imaging width W′_img below the diffraction limit σ₀ and the actual width of the defect, obtaining a correlation function therebetween, and correcting a deviation of the defect imaging width according to the correlation function, so as to obtain the defect imaging width W_img.

Beneficial effects of the present invention are as follows.

The present invention proposes a real-time traceability method of a width of a defect based on divide-and-conquer, which solves the problems that the width is difficult to be measured in real time in the defect detection process and the width of the defect close to the diffraction limit of the system is difficult to be detected. This method integrates the CTF, the extreme random tree regression model based on the multidimensional eigenvectors in the electromagnetic field simulation database of the defect scattered dark-field imaging, and the adaptive threshold segmentation algorithm to trace the width of the defect in real time. According to the method provided by the present invention, the mean square deviation of width recognition is better than 0.3 μm. The method provided by the present invention is able to not only be applied to other systems, but also construct the electromagnetic field simulation database of the defect scattered dark-field imaging which is easily combined with the future Internet of Things and big data analysis, which is conducive to promoting the development of defect detection technology in the direction of intelligence and service.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a general flow chart of a method provided by the present invention.

FIG. 2(a) shows a calibration transfer function at a magnification of 16 times.

FIG. 2(b) shows a calibration error curve at a magnification of 16 times.

FIG. 3 is a flow chart of multidimensional eigenvector analysis technology based on the electromagnetic field simulation database of defect scattered dark-field imaging.

FIG. 4 is a schematic diagram of a curved Gaussian pyramid.

FIG. 5(a) shows the comparison between the traceability width and the actual width at a magnification of 9.7 times.

FIG. 5(b) is a schematic diagram of the difference between the traceability width and the actual width at a magnification of 9.7 times.

FIG. 6(a) is a comparison diagram of the traceability width and the actual width obtained when the traceability method provided by the present invention is applied to defect dark-field images collected by the Olympus microscope dark-field imaging system.

FIG. 6(b) is a schematic diagram of the difference between the traceability width and the actual width obtained when the traceability method provided by the present invention is applied to defect dark-field images collected by the Olympus microscope dark-field imaging system.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT

The present invention is further described with accompanying drawings and specific embodiments in detail as follows.

According to FIG. 1 of the drawings, a real-time traceability method of a width of a defect based on divide-and-conquer is illustrated, which comprises steps of.

(1) estimating the width of the defect through a calibration transfer function (CTF) based on standard defect digital calibration technology, and executing a step of (2) if the width of the defect is less than 5 μm, wherein the calibration transfer function is obtained through steps of:

(1-1) according to a principle of binary optics, etching a standard defect on a fused silica glass substrate through electron beam exposure and ion beam etching technology, and obtaining a standard plate;

(1-2) measuring an actual width of the standard defect on the standard plate with a scanning electron microscope (SEM);

(1-3) collecting a dark-field image of the standard defect on the standard plate with a dark-field imaging system;

(1-4) performing a series of processing such as threshold segmentation and morphological operations on the dark-field image, and obtaining a pixel width of the standard defect; and

(1-5) taking the pixel width of the standard defect as an independent variable, and the actual width of the standard defect measured by the SEM as a dependent variable, and fitting the independent variable and the dependent variable through a least square method, so as to obtain an object-image correlation function which is the calibration transfer function,

wherein FIG. 2(a) shows a calibration transfer function at a magnification of 16 times; it is able to be seen that the calibration transfer function has a good linear law in a width range above 5 μm, and shows a nonlinear law in a width range below 5 μm; a same pixel width corresponds to multiple actual widths, leading to indistinguishable from each other;

(2) estimating the width of the defect through a multidimensional eigenvector analysis technology based on an electromagnetic field simulation database of defect scattered dark-field imaging, as shown in FIG. 3, and executing a step of (3) if the width of the defect is less than a diffraction limit of the system, wherein a regression width of the defect is obtained through steps of:

(2-1) building the electromagnetic field simulation database of defect scattered dark-field imaging, and obtaining simulation data of scattered light intensity distribution of multiple defects;

(2-2) collecting a dark-field image of a defect to be tested with the dark-field imaging system, obtaining a gray distribution along a width direction of the standard defect on the standard plate, and obtaining experimental data of the gray distribution of an ideal imaging system after deconvolution;

(2-3) performing coordinate system unity and Min-Max normalization on the simulation data obtained by the step of (2-1) and the experimental data obtained by the step of (2-2), and obtaining normalized simulation data of scattered light intensity distribution and normalized experimental data of the gray distribution, wherein the step of (2-3) specifically comprises:

(2-3-1) taking a physical size x as an abscissa for the simulation data obtained by the step of (2-1) and a pixel p as an abscissa for the experimental data obtained by the step of (2-2), unifying the simulation data obtained by the step of (2-1) and the experimental data obtained by the step of (2-2) to take the physical size x as the abscissa through formulas of

$\begin{array}{cc}p=\frac{Mx}{s_p},& \left(I\right)\end{array}$ $\begin{array}{cc}{f}_{\mathrm{res}}\left(p\right)\stackrel{p=Mx/s_p}{⟶}{f}_{\mathrm{res}}\left(x\right),& \left(\mathrm{II}\right)\end{array}$
here, a pixel size of a charge-coupled device (CCD) of the dark-field imaging system is expressed by Sp, a magnification of the dark-field imaging system is expressed by M; and

(2-3-2) performing Min-Max normalization on the simulation data obtained by the step of (2-1) and the experimental data obtained by the step of (2-2) by a formula of

$\begin{array}{cc}\{\begin{array}{c}{I}_{\mathrm{norm}}\left(x\right)=\frac{I_{\mathrm{sct}}\left(x\right)-\min \left\{I_{\mathrm{sct}}\left(x\right)\right\}}{\max \left\{I_{\mathrm{sct}}\left(x\right)\right\}-\min \left\{I_{\mathrm{sct}}\left(x\right)\right\}}\\ f_{\mathrm{norm}}\left(x\right)=\frac{f_{\mathrm{res}}\left(x\right)-\min \left\{f_{\mathrm{res}}\left(x\right)\right\}}{\max \left\{f_{\mathrm{res}}\left(x\right)\right\}-\min \left\{f_{\mathrm{res}}\left(x\right)\right\}}\end{array};& \left(\mathrm{III}\right)\end{array}$

(2-4) respectively extracting a multidimensional eigenvector of the normalized simulation data of the scattered light intensity distribution and a multidimensional eigenvector of the normalized experimental data of the gray distribution, both of which respectively comprise curved Gaussian pyramid features of the normalized simulation data and those of the normalized experimental data, wherein as shown in FIG. 4, the curved Gaussian pyramid is a series of curve data collections arranged in a pyramid shape, is derived from the same curve but with decreasing resolution, and is able to realize multi-scale representation of data; the higher the pyramid, the lower the resolution and size of the curve; the sampling points of the upper curve are half of the lower curve; an extracting method of the curved Gaussian pyramid comprises:

(2-4-1) recording distributed data as ƒ(n), wherein n represents discrete data points, ƒ_J(n) represents ƒ(n) is located at a J^th floor of a pyramid;

(2-4-2) recording a curve distribution of a j^th floor of the pyramid as ƒ_j(n), performing low-pass Gaussian filtering on ƒ_j(n), and obtaining ƒ′_j(n);

(2-4-3) downsampling ƒ′_j(n) by a factor of 2, and obtaining a curve distribution of a (j−1)^th floor of the pyramid, which is able to be regarded as discarding one data point every one data point, that is
ƒ_j−1(n)=ƒ₂(n)=ƒ_j(2n)  (IV);

(2-4-4) repeating the steps of (2-4-2) and (2-4-3) for P times, wherein j=J, J−1, . . . , J−P+1, and obtaining a P+1 layer Gaussian pyramid; and

(2-4-5) connecting data of the P+1 layer Gaussian pyramid with each other in series, and obtaining the multidimensional eigenvectors;

(2-5) taking the multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution in the step of (2-4) as a sample set, training through an extreme random tree regression algorithm with supervised learning in machine learning theory, and obtaining a regression model of a width of any defect; and

(2-6) taking the multidimensional eigenvectors of the normalized experimental data of the gray distribution in the step of (2-4) as a test set, inputting the regression model obtained by the step of (2-5), estimating the width of the defect to be tested, and obtaining the regression width of the defect to be tested,

wherein the step of (2-1) building the electromagnetic field simulation database of defect scattered dark-field imaging, and obtaining simulation data of scattered light intensity distribution of multiple defects are specifically as follows:

(2-1-1) building an electromagnetic field simulation model of defect scattered dark-field imaging with FDTD (finite-difference time-domain) simulation software;

(2-1-2) building a simulation model of common defects;

(2-1-3) performing numerical calculations of electromagnetic fields with the FDTD simulation software, and obtaining the scattered light intensity distribution of the defects;

(2-1-4) performing near field—far field projection, obtaining light intensity distribution of the defects on a far-field image plane of an ideal optical imaging system, and recording as I_sct(x); and

(2-1-5) changing a defect shape and a defect size, continuously repeating the steps of (2-1-3) to (2-1-4), obtaining the scattered light intensity distribution of the defects with different shapes and sizes, obtaining a series of simulation data, wherein the series of simulation data comprise the shape of the defects, the width of the defects and the light intensity distribution I_sct(x) of the defects, so as to form the simulation data of the scattered light intensity distribution of the defects.

The step of (2-2) that collecting the dark-field image of the defect to be tested with the dark-field imaging system, obtaining a gray distribution along a width direction of the standard defect on the standard plate, and obtaining experimental data of the gray distribution of an ideal imaging system after deconvolution are specifically as follows:

(2-2-1) collecting the dark-field image of the defect to be tested with the dark-field imaging system, and extracting the gray distribution ƒ_exp(p) along the width direction of the dark-field image of the defect to be tested;

(2-2-2) measuring a point spread function (PSF) of the dark-field imaging system through a checkerboard diagram, wherein since a PSF of an actual optical system tends to a form of Gaussian function, an expression of the PSF of the dark-field imaging system is obtained through Gaussian function fitting in a measurement process; and

(2-2-3) taking the PSF obtained by the step of (2-2-2) as an initial value, deconvolving the gray distribution ƒ_exp(p) obtained by the step of (2-2-1), and obtaining the gray distribution ƒ_res(p) of the ideal imaging system after deconvolution.

The step of (2-2-1) that extracting the gray distribution ƒ_exp(p) along the width direction of the dark-field image of the defect to be tested is specifically as follows.

1) performing Otsu threshold segmentation on an original dark-field image, selecting an elongated structural element for opening operations to connect line features of fractures and holes while removing noise, and then extracting long line features which are defect characteristics;

2) performing linear fitting on pixel coordinates (u_i, v_i) of pixels that form the line features through the least square method, and obtaining a straight line expression of av+b=u, wherein both a and b are parameters;

3) obtaining a positive angle γ between the defect and v axis through the parameter a, wherein γ=arctan (α), and rotate the original dark-field image around an upper left corner for −γ, so that the defect is rotated to a vertical direction;

4) calculating a gray histogram of the original dark-field image, wherein a peak value at low gray denotes a gray level of image background, filling a blank area that appears after rotation with the gray of image background, and obtaining a filled image to maintain an image integrity;

5) taking a column average of the filled image as a gray distribution ƒ₀(p) of the defect along the width direction, which is unimodal or bimodal, taking a vertical straight line where a peak abscissa p₀ of the unimodal distribution is a symmetry axis, assuming that a data width is 2N+1, and intercepting data points between p₀−N and p₀+N as a gray distribution ƒ_N(p):

6) averaging after making the gray distribution ƒ_N(p) axisymmetric, and obtaining an axisymmetric distribution ƒ_sym(p) which is expressed by a formula of

$\begin{array}{cc}{f}_{\mathrm{sym}}\left(p\right)=\frac{1}{2}\left[f_N\left(p\right)+f_N\left(2p_0-p\right)\right],pϵ\left[p_0-N,p_0+N\right],& \left(V\right)\end{array}$
moving coordinate axes, taking an abscissa of the symmetry axis as a coordinate origin, and obtaining a gray distribution ƒ_exp(p) of the dark-field image of the defect along the width direction, which is expressed by a formula of
ƒ_exp(p)=ƒ_sym(p−p₀),p∈[−N,N]  (VI).

wherein if the regression width W_regrs is less than the diffraction limit of the system, the dark-field image of the defect is severely degraded, an effect after deconvolution is poor; the grayscale distribution in this width range is very close after normalization, the regression model is unable to distinguish different widths in this width range. Therefore, the width traceability in this width range is performed through the method described in the step of (3).

The step of (3) comprises processing the dark-field image according to adaptive threshold segmentation method, and obtaining the defect imaging width W_img, which is specifically as follows.

(3-1) performing threshold segmentation on the dark-field image through Otsu algorithm;

(3-2) extracting an area A and a length l of the defect, and obtaining a pixel width W_p of the defect, wherein w_p=A/l;

(3-3) through the pixel size S_p of the CCD, and the magnification A of the dark-field imaging system, obtaining a simple imaging width W′_img which is expressed by a formula of

$\begin{array}{cc}{w}_{\mathrm{img}}^{\prime }=\frac{s_p}{M}{w}_p;& \left(\mathrm{VII}\right)\end{array}$
and
(3-4) performing linear fitting on the simple imaging width W′_img below the diffraction limit σ₀ and the actual width of the defect, obtaining a correlation function therebetween, and correcting a deviation of the defect imaging width according to the correlation function, so as to obtain the defect imaging width W_img.

First Embodiment

The present invention is able to effectively trace defects whose width is close to the diffraction limit of the dark-field imaging system. Combining the regression width W_regrs and the imaging width W_img, taking the diffraction limit of the dark-field imaging system as the boundary, the dark-field imaging system adopts white light for illumination, has the center wavelength λ of 550 nm, the numerical aperture (NA) of 0.15 for high magnification (9.7 times) detection, so that the diffraction limit of the dark-field imaging system at this time is obtained, namely,

${\sigma }_0=\frac{0.5\lambda }{\mathrm{NA}}\approx 1.83\mathrm{\mu m}.$
The traceability method provided by the present invention is used to trace 6 defects with theoretical widths of 0.5, 1, 2, 3, 4 and 5 μm on 8 standard plates with depths of 0.028, 0.053, 0.128, 0.149, 0.205, 0.257, 0.503 and 1.155 μm, respectively, and accordingly, 6×8=48 traceability results of the width of the defect are obtained. FIGS. 5(a) and 5(b) show the results of the real-time traceability method of the width of the defect based on divide-and-conquer. In FIG. 5(a), the abscissa is the sample number, discrete data points represent the actual width, the solid line indicates the traceability width. FIG. 5(b) shows the difference between the traceability width and the actual width is in the range of ±0.8 μm.

Second Embodiment

The present invention is able to effectively trace defects in different dark-field imaging systems, as long as relevant parameters thereof are replaced. The relevant parameters are:

1) the magnification and the pixel size of the CCD, wherein the magnification and the pixel size of the CCD are used in all places where an imaging coordinate system and a pixel coordinate system need to be converted, and the magnification and the pixel size of the CCD need to be replaced during a conversion process;

2) the numerical aperture and the diffraction limit, wherein different detection systems have different numerical apertures (NAs), different diffraction limits σ₀ are calculated for being replaced while using a decision model library through a formula of

$\begin{array}{cc}{\sigma }_0=\frac{0.5\lambda }{\mathrm{NA}};,& \left(\mathrm{VIII}\right)\end{array}$

here, λ is a wavelength of a light source of a corresponding dark-field detection system;

3) the point spread function, wherein different detection systems have different PSFs, which are calibrated again through the checkerboard diagram and do not meet requirements of rapid testing; since the deconvolution does not need to be based on prior PSF knowledge, only an initial estimation of the PSF is required, a PSF calculation method of the other dark-field imaging systems is expressed by a formula of

$\begin{array}{cc}{h}^{\prime }\left(p\right)=\{\begin{array}{c}K\exp \left(-\frac{p^2}{2{\sigma }^{\mathrm{\prime 2}}}\right)\\ 0,\mathrm{others}\end{array},pϵD,& \left(\mathrm{IX}\right)\end{array}$

here

${\sigma }^{\prime }=\frac{\sigma M}{M^{\prime }},$
σ is a standard error of an original PSF, M is a magnification of an original dark-field imaging system, M′ is a magnification of the other dark-field imaging systems, K is a constant.

In the dark-field imaging mode of the OlympusBX51M metallurgical microscope, a 10× objective lens is used, NA^Olym=0.3, the pixel size of the CCD s_p^Olym=3.45 μm, and there is a 0.63× imaging lens in front of the CCD. The magnification, the diffraction limit, and the point spread function of the traceability method are replaced as follows:

1) Magnification:
M^Olym=10×0.63=6.3  (X).

2) Diffraction Limit:

$\begin{array}{cc}{\sigma }_0^{\mathrm{Olym}}=\frac{0.5\lambda }{{\mathrm{NA}}^{\mathrm{Olym}}}\approx 0.9\mathrm{\mu m}.& \left(\mathrm{XI}\right)\end{array}$

3) Point Spread Function:

$\begin{array}{cc}{h}^{\mathrm{Olym}}\left(p\right)=10.71\exp \left[-{\left(\frac{p}{5.503⨯\frac{9.7}{6.3}}\right)}^2\right]=10.71\exp \left[-{\left(\frac{p}{8.525}\right)}^2\right].& \left(\mathrm{XII}\right)\end{array}$
The dark-field images of defects with widths of 0.5, 1, 2, 3, 4, and 5 μm on the standard plate with depths of 0.128, 0.149, 0.205, 0.257, 0.503 μm are collected, respectively. The model after parameter replacement is used to identify the widths. FIGS. 6(a) and 6(b) show the results of the real-time traceability method of the width of the defect based on the divide-and-conquer. In FIG. 6(a), the abscissa is the sample number, discrete data points represent the actual width, the solid line indicates the traceability width. FIG. 6(b) shows the difference between the traceability width and the actual width is in the range of +0.65 μm.

Claims

1. A real-time traceability method of a width of a defect based on divide-and-conquer, the real-time traceability method comprising steps of:

(1) estimating the width of the defect through a calibration transfer function (CTF) based on standard defect digital calibration technology, recording the width of the defect estimated through the CTF as WCTF, wherein if WCTF is greater than 5 μm, which means that the width of the defect complies with geometric imaging law and is located in a linear segment of the CTF, then WCTF acts as an accurately estimated value of an actual width of the defect; if WCTF is less than or equal to 5 μm, which means that the width of the defect is smaller, diffraction broadening occurs during imaging, the width of the defect no longer complies with the geometric imaging law, and WCTF acting as the estimated value of the actual width of the defect has a large error, then at this time, a step of (2) is executed;

(2) estimating the width of the defect through a multidimensional eigenvector analysis technology based on an electromagnetic field simulation database of defect scattered dark-field imaging; recording the width of the defect estimated through the multidimensional eigenvector analysis technology as a regression width Wregrs, wherein if Wregrs is greater than a diffraction limit σ0 of a dark-field imaging system, then the regression width Wregrs acts as the accurately estimated value of the actual width of the detect; if Wregrs is less than or equal to the diffraction limit σ0, the regression width Wregrs is unable to distinguish different widths in a width range which is less than or equal to the diffraction limit σ0, then at this time, a step of (3) is executed; and

(3) processing a dark-field image and extracting the width of the defect through an adaptive threshold segmentation method, recording the width of the defect extracted through the adaptive threshold segmentation method as an imaging width Wimg, wherein if the imaging width Wimg is greater than the diffraction limit σ0, the imaging width Wimg is unable to distinguish different widths, then the regression width Wregrs still acts as the estimated value of the actual width of the defect; if the imaging width Wimg is less than or equal to the diffraction limit σ0, then the imaging width Wimg acts as the estimated value of the actual width of the defect.

2. The real-time traceability method of the width of the defect based on divide-and-conquer according to claim 1, wherein the step of (2) that, through the multidimensional eigenvector analysis technology based on the electromagnetic field simulation database of defect scattered dark-field imaging, the regression width is obtained, specifically comprises:

(2-1) building the electromagnetic field simulation database of defect scattered dark-field imaging, and obtaining simulation data of scattered light intensity distribution of multiple defects on a far-field CCD (charged-couple device) image surface;

(2-2) collecting a dark-field image of a defect to be tested with the dark-field imaging system, obtaining a gray distribution along a width direction of the standard defect on the standard plate, and obtaining experimental data of the gray distribution of an ideal imaging system after deconvolution;

(2-3) performing coordinate system unity and Min-Max normalization on the simulation data obtained by the step of (2-1) and the experimental data Obtained by the step of (2-2), and obtaining normalized simulation data of scattered light intensity distribution and normalized experimental data of the gray distribution;

(2-4) respectively extracting multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution and multidimensional eigenvectors of the normalized experimental data of the gray distribution:

(2-5) taking the multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution in the step of (2-4) as a sample set, training through an extreme random tree regression algorithm with supervised learning in machine learning theory, and obtaining a regression model of a width of any defect; and

(2-6) taking the multidimensional eigenvectors of the normalized experimental data of the gray distribution in the step of (2-4) as a test set, inputting the regression model obtained by the step of (2-5), estimating the width of the defect to be tested, and obtaining the regression width Wregrs of the defect to be tested.

3. The real-time traceability method of the width of the defect based on divide-and-conquer according to claim 2, wherein the step of (2-2) that, collecting the dark-field image of the defect to be tested with the dark-field imaging system, obtaining the gray distribution along the width direction of the standard defect on the standard plate, and obtaining experimental data of the gray distribution of the ideal imaging system after deconvolution, specifically comprises:

(2-2-1) collecting the dark-field image of the defect to be tested with the dark-field imaging system, and extracting the gray distribution ƒexp(p) along the width direction of the dark-field image of the defect to be tested;

(2-2-2) measuring a point spread function (PSF) of the dark-field imaging system through a checkerboard diagram, wherein since a PSF of an actual optical system tends to a form of Gaussian function, an expression of the PSF of the dark-field imaging system is Obtained through Gaussian function fitting in a measurement process; and

(2-2-3) taking the PST Obtained by the step of (2-2-2) as an initial value, deconvolving the gray distribution ƒexp(p) obtained by the step of (2-2-1), and obtaining the gray distribution ƒres(p) of the ideal imaging system after deconvolution.

4. The real-time traceability method of the width of the defect based on divide-and-conquer according to claim 3, wherein the step of (2-2-1) that, extracting the gray distribution ƒexp(p) along the width direction of the dark-field image of the defect to be tested, specifically comprises: f sym ( p ) = 1 2 [ f N ( p ) + f N ( 2 ⁢ p 0 - p ) ], p ⁢ ϵ [ p 0 - N, p 0 + N ], ( I )

1) performing Otsu threshold segmentation on an original dark-field image, selecting an elongated structural element for opening operations to connect line features of fractures and holes while removing noise, and then extracting long line features which are defect characteristics;

2) performing linear fitting on pixel coordinates (ui, vi) of pixels that form the line features through the least square method, and obtaining a straight line expression of av+b=u, wherein both a and b are parameters;

3) obtaining a positive angle γ between the defect and v axis through the parameter a, wherein γ=arctan (α), and rotate the original dark-field image around an upper left corner for −γ, so that the defect is rotated to a vertical direction;

4) calculating a gray histogram of the original dark-field image, wherein a peak value at low gray denotes a gray level of image background, idling a blank area that appears after rotation with the gray of image background, and obtaining a filled image to maintain an image integrity;

5) taking a column average of the filled image as a gray distribution ƒ0(p) of the detect along the width direction, which is unimodal or bimodal, taking a vertical straight line where a peak abscissa p0 of the unimodal distribution is a symmetry axis, assuming that a data width is 2N+1, and intercepting data points between p0−N and p0+N as a gray distribution ƒN(p); and

6) averaging after making the gray distribution ƒN(p) axisymmetric, and obtaining an axisymmetric distribution ƒsym(p) which is expressed by a formula of

moving coordinate axes, taking an abscissa of the symmetry axis as a coordinate origin, and obtaining a gray distribution ƒexp(p) of the dark-field image of the defect along the width direction, which is expressed by a formula of ƒexp(p)=ƒsym(p−p0),p∈[−N,N]   (II).

5. The real-time traceability method of the width of the defect based on divide-and-conquer according to claim 2, wherein in the step of (2-4), the multidimensional eigenvectors of the normalized simulation data of the scattered light intensity distribution and the multidimensional eigenvectors of the normalized experimental data of the gray distribution, respectively comprise curved Gaussian pyramid features of the normalized simulation data and those of the normalized experimental data, an extracting method of the multidimensional eigenvectors comprises:

(2-4-1) recording distributed data as ƒ(n), wherein n represents discrete data points, ƒJ(n) represents ƒ(n) is located at a Jth floor of a pyramid;

(2-4-2) recording a curve distribution of a jth floor of the pyramid as ƒj(n), performing low-pass Gaussian filtering on ƒJ(n), and obtaining ƒ′j(n);

(2-4-3) downsampling ƒ′j(n) by a factor of 2, and obtaining a curve distribution of a (j−1)th floor of the pyramid, which is able to be regarded as discarding one data point every one data point and is expressed by a formula of ƒj−1(n)=ƒ2(n)=ƒj(2n)   (III)

(2-4-4) repeating the steps of (2-4-2) and (2-4-3) for P times, wherein j=J, J−1,..., J−P+1, and obtaining a P+1 layer Gaussian pyramid; and

(2-4-5) connecting data of the P+1 layer Gaussian pyramid with each other in series, and obtaining the multidimensional eigenvectors.

6. The real-time traceability method of the width of the defect based on divide-and-conquer according to claim 5, wherein the step of (3) that, processing the dark-field image according to adaptive threshold segmentation method, and obtaining the defect imaging width Wimg, specifically comprises: w img ′ = s p M ⁢ w p; ( IV )

through the pixel size Sp of the CCD, and the magnification M of the dark-field imaging system, obtaining a simple imaging width trim w′img which is expressed by a formula of

 and

performing linear fitting on the simple imaging width Wimg below the diffraction limit σ0 and the actual width of the defect, obtaining a correlation function therebetween, and correcting a deviation of the defect imaging width according to the correlation function, so as to obtain the defect imaging width Wimg.

7. The real-time traceability method of the width of the defect based on divide-and-conquer according to claim 6, wherein the present invention has portability for defect detection under other dark-field imaging systems, and only needs to obtain relevant parameters of the other dark-field imaging systems for replacement, which specifically comprises: h ′ ( p ) = { K ⁢ exp ⁡ ( - p 2 2 ⁢ σ ′2 ) 0, others, p ⁢ ϵ ⁢ D, ( V ) σ ′ = σ ⁢ M M ′,

obtaining the magnification and the pixel size of the CCD, wherein the magnification and the pixel size of the CCD are used in all places where an imaging coordinate system and a pixel coordinate system need to be converted, and the magnification and the pixel size of the CCD need to be replaced during a conversion process;

obtaining the numerical aperture and calculating the diffraction limit, wherein different detection systems have different numerical apertures (NAs), different diffraction limits σ0 are calculated for being replaced while using a decision model library;

calculating the point spread function, wherein different detection systems have different PSFs, which are calibrated again through the checkerboard diagram and do not meet requirements of rapid testing; since the deconvolution does not need to be based on prior PSF knowledge, only an initial estimation of the PSF is required, a PSF calculation method of the other dark-field imaging systems is expressed by a formula of

here,

 σ is a standard error of an original PST, M is a magnification of an original dark-field imaging system, M′ is a magnification of the other dark-field imaging systems, K is a constant.